Tray detecting and positioning method based on radio frequency identification and sensor fusion

Abstract

The invention discloses a tray detecting and positioning method based on radio frequency identification and sensor fusion, which comprises the following steps: deploying an RFID tag matrix and multiple types of sensors on a tray to realize data space-time synchronous acquisition; fusing multi-modal data by using an improved D-S evidence theory and the like, and adaptively adjusting the importance of each modal through a dynamic weight factor; a three-dimensional space and attitude constraint equation is constructed, a Levenberg-Marquardt algorithm is used for solving, the accurate coordinates and attitude of the tray are output, and accurate detection and positioning are achieved. According to the invention, through RFID and multi-sensor data fusion and innovative algorithm processing, all-directional sensing of static and dynamic information of the tray is realized; the system can accurately position trays, adapt to a complex environment, effectively process abnormal conditions, optimize energy consumption, improve multi-tray cooperation efficiency, and improve logistics management intelligent level and operation efficiency.

Description

Technical Field

[0001] This invention relates to the field of pallet positioning technology, and in particular to a pallet detection and positioning method based on radio frequency identification and sensor fusion. Background Technology

[0002] In the field of modern logistics warehousing and supply chain management, pallets serve as the basic unit for carrying and transferring goods. Their accurate detection and positioning are crucial for achieving automated operations and improving operational efficiency. Traditional pallet management relies heavily on manual visual inspection and barcode scanning, which is not only inefficient but also susceptible to factors such as ambient light and barcode damage, making it difficult to meet the demands of large-scale, high-frequency goods handling. With the intelligent upgrading of the logistics industry, pallet management systems based on Radio Frequency Identification (RFID) technology are gradually becoming more widespread. While this technology enables contactless pallet identification and information reading, relying solely on RFID technology makes it difficult to obtain dynamic information such as the pallet's real-time posture, movement status, and environmental parameters. In complex warehousing environments, it suffers from insufficient positioning accuracy and weak multi-pallet differentiation capabilities.

[0003] Meanwhile, sensor-based positioning technologies, such as inertial sensors, ultrasonic sensors, and Bluetooth sensors, can collect data on pallet movement and the environment. However, single sensors suffer from limitations such as limited data dimensions and susceptibility to interference. For example, inertial sensors accumulate errors over time, and ultrasonic sensors exhibit significant positioning deviations in environments with pronounced multipath effects. Some existing solutions attempt to combine RFID with single-type sensors, but these often employ simple data stitching methods, lacking deep fusion processing of multi-source heterogeneous data. This fails to fully exploit the potential correlations between data points, making it difficult to achieve comprehensive perception and accurate positioning of the pallet's status.

[0004] Furthermore, existing pallet detection and positioning technologies have significant shortcomings in terms of adaptability to dynamic environments and handling of abnormal conditions. In the warehousing environment, factors such as changes in shelf layout, obstruction by stacked goods, and electromagnetic interference frequently change, making it difficult for traditional methods to dynamically adjust positioning strategies according to environmental changes. When sensor data is abnormal or RFID signals are lost, the system lacks effective fault tolerance mechanisms and data repair capabilities, leading to positioning failures or result deviations, which seriously affect the continuity and reliability of logistics operations. Therefore, there is an urgent need for a pallet detection and positioning technology that can integrate the advantages of multi-source data, adapt to complex environmental changes, and possess strong robustness. Summary of the Invention

[0005] To address the problems existing in the prior art, the present invention aims to provide a pallet detection and positioning method based on radio frequency identification and sensor fusion, which enables comprehensive perception of both static and dynamic information of pallets; it can accurately locate pallets, adapt to complex environments, effectively handle abnormal situations, optimize energy consumption, improve the efficiency of multi-pallet collaboration, and enhance the level of intelligent logistics management and operational efficiency.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A pallet detection and positioning method based on radio frequency identification and sensor fusion includes the following steps:

[0008] 1) Multi-source heterogeneous data acquisition and spatiotemporal synchronization steps: Install an array of UHF RFID tags on the surface of the pallet, with each tag storing feature information of different areas; install a MEMS six-axis sensor in the pallet; install a Bluetooth ranging sensor array in the warehouse; achieve time synchronization of all sensor data through the IEEE1588 time protocol to construct a spatiotemporally consistent data cube;

[0009] 2) Multimodal feature fusion and dynamic weighted localization steps: Data fusion is performed using DS evidence theory, through formulas... Calculate the basic probability distribution after fusion; where the conflict coefficient is... m1(A i ) and m2(A j Proposition A represents RFID and sensor data pairs, respectively. i and A j Basic probability allocation; introduction of dynamic weighting factors Where λ = 0.5 is the forgetting factor, x t The feature value of the sensor data at time t is used to adaptively adjust the modal weights based on the rate of change of the sensor data.

[0010] 3) Multidimensional constraint localization solution steps: Constructing three-dimensional spatial constraint equations Where (x,y,z) are the three-dimensional coordinates of the pallet to be positioned, (x,y,z) i y i , z i Let be the three-dimensional coordinates of the i-th reference point, and w be the coordinates of the i-th reference point. i Let d be the weight coefficient for the i-th reference point. i f(θ,φ,ψ) represents the distance measurement from the pallet to be positioned to the i-th reference point; f(θ,φ,ψ) represents the approximate pallet orientation d. i The beam functions, θ, φ, and ψ, represent the roll angle, pitch angle, and yaw angle of the pallet, respectively. The Levenberg-Marquardt algorithm is used to iteratively solve the nonlinear equations, with a convergence threshold set to 10. -6 .

[0011] Furthermore, it also includes:

[0012] Dynamic environment perception and adaptive positioning steps: Real-time classification of environment types using machine learning algorithms, and dynamic adjustment of positioning parameters based on environmental characteristics; in the shelving area, an shelving reflection signal propagation model is introduced. Compensation for the multipath effect; where d eff The effective propagation distance is d, where d is the direct propagation distance and k is k. i Let Δd be the reflection coefficient of the i-th reflection path. i The additional path length for the i-th reflection path;

[0013] Anomaly detection and robustness enhancement steps: An anomaly detection algorithm based on phase space reconstruction is designed. The attractor trajectory of the reconstructed phase space from sensor data is calculated. An anomaly handling mechanism is triggered when the trajectory deviation exceeds a threshold ∈ = 1.5σ. A robust Kalman filter algorithm is employed. in This is the optimal state estimate at time k. Let K be the predicted state at time k. k Let z be the gain matrix. k Let H be the observation value at time k, and H be the observation matrix; K be the gain matrix. k =P k|k-1 H T HP k|k-1 H T +R+λI) -1 λ is the robustness factor, P k|k-1 To predict the covariance matrix, R is the observation noise covariance matrix, and I is the identity matrix.

[0014] Furthermore, in the multimodal feature fusion and dynamic weighted localization steps, a deep learning model is used for feature fusion, and the importance weights of each modality are automatically learned through an attention mechanism, as shown in the formula: Among them, h i Let W1 and W2 be the feature vectors of the i-th mode, and W1 and W2 be the learnable weight matrices, α i Let be the attention weight for the i-th modality.

[0015] Furthermore, in the multi-dimensional constraint positioning solution step, inertial navigation-assisted positioning is introduced, using the formula... Calculate the displacement increment; where Δp is the displacement increment, v(t) is the velocity at time t, and a(τ) is the acceleration at time τ; inertial data and RFID data are fused using complementary filtering, and the fusion formula is as follows: in The position estimate after fusion at time t. For the position estimate at time t-1, To integrate weights, These are the variances of RFID and inertial measurement, respectively. The position data of the inertial measurement unit at time t.

[0016] Furthermore, in the dynamic environment perception and adaptive localization step, an environmental fingerprint database is constructed using a Gaussian mixture model. Perform environment matching; where p(x) is the probability density function of environment feature x, K is the number of Gaussian distributions, and π k Let be the mixing coefficient of the k-th Gaussian distribution. The mean is μ k The covariance matrix is ​​∑ k Gaussian distribution; Kullback-Leibler divergence is used. Calculate the similarity between the current environment and the fingerprint database, where p(x) and q(x) are two probability distribution functions.

[0017] Furthermore, in the abnormal state detection and robustness enhancement steps, a temporal anomaly scoring mechanism is designed, using a formula... Calculate the outlier score; where S t Let w be the outlier score at time t, N be the sliding window size, and w be the outlier score at time t. i =e -i / τ The time decay weight is τ = 5, which is the time constant, and x is the time decay weight. t-i For the actual data at time ti, For the predicted data at time ti; when S t An alarm is triggered when the threshold is exceeded for three consecutive times.

[0018] Furthermore, it also includes:

[0019] Energy consumption optimization management steps: Establish sensor energy consumption model Among them, E total For total energy consumption, E i Let be the energy consumption of the i-th sensor, and α = 0.01 be the nonlinear coefficient; a particle swarm optimization algorithm is used. Dynamically adjust the sensor sampling frequency; where Let be the velocity vector of particle i at time t+1, and w = 0.7 be the inertial weight. Let be the velocity vector of particle i at time t, c1 = c2 = 1.4 be the acceleration constant, r1 and r2 be random numbers between [0,1], and p i Let be the optimal position of particle i. Let p be the position of particle i at time t. g This is the globally optimal position.

[0020] Multi-pallet collaborative positioning steps: When each pallet enters the positioning area, construct the relative position constraint equations between the pallets. Where p i p j Let d be the coordinate vectors of tray i and tray j, respectively. ij The relative distance measurement between tray i and tray j, ∈ ij To mitigate measurement error, a graph optimization algorithm (min) is employed.x ∑ i，j∈ε e ij (x) T Ω ij e ij (x) Solve for the global optimal position; where X is the set of position variables for all pallets, ε is the edge set of relative position constraints between pallets, and e ij (x) is the error function of the relative positions of tray i and tray j, Ω ij This is an information matrix.

[0021] Furthermore, in the multi-source heterogeneous data acquisition and spatiotemporal synchronization step, compressed sensing technology is used to acquire sensor data, and data compression is achieved through the formula y = Φx; where x is the original signal vector, Φ is the measurement matrix, and y is the compressed data vector; the reconstruction algorithm adopts the orthogonal matching pursuit (OMP) algorithm, with a number of iterations... μ = 1.2 is the overcompleteness coefficient, and σ0(x) is the sparsity of the original signal x.

[0022] Furthermore, in the multi-dimensional constraint positioning solution step, multi-frequency RFID signal fusion technology is adopted, through the formula... Calculate the distance; where d is the distance from the tray to be positioned to the RFID reader, c is the speed of light, f is the RFID signal operating frequency, and Phase is the RFID signal phase difference; construct a weighted least squares problem by combining multi-frequency measurements. Where x is the coordinate of the pallet to be positioned. The weight of the measurement value at the i-th frequency point. Let p be the variance of the measurement at the i-th frequency point, di be the distance measured at the i-th frequency point, and p be the variance of the measurement at the i-th frequency point. i Let be the coordinates of the i-th reference point.

[0023] Compared with existing technologies, the beneficial effects of this invention are:

[0024] At the data acquisition level, an RFID tag matrix and a multi-type sensor array work together to achieve comprehensive perception of the pallet's static information and dynamic status. Compared with traditional single technology solutions, the data acquisition dimensions increase by more than three times, effectively solving the problem of missing information.

[0025] In the data processing and positioning stages, innovative multimodal feature fusion algorithms and dynamic weighting mechanisms deeply mine the correlations between multi-source data. Combined with multidimensional constrained positioning models and optimization algorithms, pallet positioning accuracy is improved to the centimeter level, more than 60% higher than traditional fusion methods. This effectively meets the high-precision operation requirements of automated warehousing equipment docking and precise sorting. Simultaneously, through dynamic environmental perception and adaptive positioning strategies, the system can identify different warehousing environments in real time and automatically adjust parameters. In complex scenarios such as dense shelving, the positioning success rate is increased to 98%, significantly enhancing environmental adaptability.

[0026] Anomaly detection and robustness enhancement technologies can quickly identify data anomalies and positioning failures, and use algorithms such as robust Kalman filtering to achieve data repair and positioning continuation, ensuring stable system operation under complex conditions. Furthermore, energy optimization management and multi-pallet collaborative positioning functions not only reduce the energy consumption of sensor nodes by more than 30%, but also improve the efficiency of simultaneous multi-pallet positioning by 40%, effectively reducing operating costs and improving logistics efficiency. This method comprehensively overcomes the limitations of existing technologies, providing efficient and reliable technical support for the development of smart logistics. Attached Figure Description

[0027] Figure 1 This is a schematic block diagram of a pallet detection and positioning method based on radio frequency identification and sensor fusion proposed in this invention;

[0028] Figure 2 This is a schematic diagram comparing the positioning accuracy of different pallet detection and positioning methods based on radio frequency identification and sensor fusion proposed in this invention;

[0029] Figure 3 This is a schematic diagram illustrating the success rate of a pallet detection and positioning method based on radio frequency identification and sensor fusion proposed in this invention in complex environments. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0032] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. Furthermore, the terms "installed," "connected," and "linked" should be interpreted broadly; for example, they may refer to a fixed connection, a detachable connection, or an integral connection; they may refer to a mechanical connection or an electrical connection; they may refer to a direct connection or an indirect connection through an intermediate medium; and they may refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. The invention will now be described in further detail with reference to the accompanying drawings.

[0033] Reference Figures 1 to 3 : A specific implementation of a pallet detection and positioning method based on radio frequency identification and sensor fusion.

[0034] Pallet dynamic positioning technology solution for intelligent logistics warehousing scenarios

[0035] Specific implementation steps

[0036] 1. Collaborative acquisition and spatiotemporal synchronization of multi-source heterogeneous data

[0037] ThingMagicM6e UHF RFID tags, operating at 922.5MHz, were deployed in a 3×3 array on the pallet surface. Each tag stores static information such as the pallet ID and weight limit, with a measured read / write distance of 3.5 meters. Bosch BNO055 six-axis sensors were installed at each of the four corners of the pallet, acquiring triaxial acceleration at a sampling frequency of 100Hz (range ±16g, accuracy 0.01m / s²). 2 A three-axis gyroscope (range ±2000° / s, accuracy 0.01° / s) is used to capture dynamic data of pallet translation and rotation. Twelve Bluetooth sensors are evenly distributed at the same height as the columns within the warehouse, forming a rectangular monitoring network with sides of 10 meters. All devices achieve nanosecond-level time synchronization via the IEEE 1588 protocol. The RFID reader polls tags at a 50ms cycle, combined with the time-slotted ALOHA algorithm to resolve signal conflicts. In actual testing, the tag recognition success rate reached 99.7%. Sensor data is stored in a circular buffer according to timestamps, constructing a spatiotemporally consistent data cube.

[0038] 2. Dynamic weight-driven multimodal feature fusion mechanism

[0039] The innovative fusion algorithm dynamically adjusts the weights of each modality based on the rate of change of sensor data. After converting RFID signal strength (RSSI) into a distance estimate using an empirical formula, it is fused with Bluetooth ranging data using an improved DS evidence theory. The core lies in introducing a dynamic weighting factor with a forgetting factor λ = 0.5: when the pallet's moving speed exceeds 0.5 m / s, by calculating the absolute value of the difference between adjacent moments in the sensor data, the inertial sensor weight is automatically increased from the default 0.4 to 0.65, while the RFID weight is correspondingly reduced to 0.35, effectively suppressing positioning drift in dynamic scenarios. This mechanism achieves a smooth weight transition through an exponential decay function, reducing fusion data error by more than 70% compared to traditional static weighting schemes. Simultaneously, a 3-layer fully connected deep learning model is used to further extract features, automatically learning the importance of each modality through an attention mechanism, improving fusion accuracy by 28% in complex scenarios such as dense shelving.

[0040] 3. Localization solution with joint constraints of three-dimensional space and attitude

[0041] A multi-dimensional constraint model incorporating spatial coordinates and attitude parameters was constructed: The coordinates of four reference points within the warehouse were accurately calibrated using a total station. Spatial constraint equations were established based on the inverse distance weighting principle, with weights decreasing quadratically with increasing distance. Quaternions were innovatively introduced to represent the pallet attitude, incorporating roll angle, pitch angle, and yaw angle into the positioning solution. Quaternion normalization constraints ensured the physical rationality of the attitude calculation. The Levenberg-Marquardt algorithm was used to iteratively solve the nonlinear equations. The initial damping factor was set to 0.01, and it was automatically reduced as the error function decreased to improve accuracy. The convergence threshold was set to 10. -6 After introducing inertial navigation-assisted positioning, complementary filtering is used to fuse RFID and inertial data: in the static state, the weight is tilted towards RFID (weight ratio 0.8), and in dynamic motion, the continuity is maintained by relying on inertial data, ultimately achieving a three-dimensional positioning accuracy of 8cm, which is 76.6% higher than the traditional RFID positioning scheme.

[0042] 4. Adaptive localization strategy based on environmental fingerprint database

[0043] The environment is classified using a random forest algorithm, extracting 15-dimensional features such as RSSI standard deviation and Bluetooth phase difference. The model is trained using 2000 samples (500 in open areas, 800 in shelving areas, and 700 in aisle areas), achieving a classification accuracy of 96%. When a metal shelving environment is detected, a multipath compensation model is automatically activated: based on pre-calibrated reflection coefficients (0.7 for metal surfaces, 0.3 for wooden shelves) and a ray tracing algorithm, the additional path length is calculated to correct the direct propagation distance. Simultaneously, an environmental fingerprint database containing five Gaussian distributions is constructed. By calculating the similarity between the current environment and the fingerprint database in real time (using the Kullback-Leibler divergence metric), the RFID signal threshold and sensor sampling frequency are dynamically adjusted. For example, in densely shelving areas, the RFID signal threshold is increased from -70dBm to -65dBm, and the Bluetooth sampling frequency is increased from 50Hz to 80Hz, improving the positioning success rate from 62% to 98% in this scenario.

[0044] 5. Robust handling of anomalies driven by phase space reconstruction

[0045] Phase space reconstruction is performed on accelerometer data. A trajectory model is constructed by embedding dimension m=5 and time delay τ=10. An anomaly alarm is triggered when the Euclidean distance between the measured trajectory and the standard trajectory exceeds 1.8 times the standard deviation for three consecutive points. A Kalman filter algorithm with a robustness factor λ=0.15 is used. When anomalies such as RFID signal loss are detected, the robustness factor is automatically increased to suppress the influence of outliers. A time-series anomaly scoring mechanism is designed, which calculates the score based on the deviation between historical data and predicted data within a sliding window (time decay weight wi=e). -i / 5 When the score continuously exceeds the standard, a linear regression model is activated to correct outliers. The anomaly handling time is reduced by 68.4% compared to the traditional method, ensuring that the system can maintain positioning continuity even when the sensor fails. Performance verification is shown in Table 1.

[0046] Table 1

[0047]

[0048]

[0049] Specific implementation steps of the multi-pallet collaborative positioning and energy consumption optimization scheme for power material warehousing:

[0050] 1. Multi-source data acquisition and compression under wide temperature range

[0051] A 5×5 array of ImpinjMonza5 UHF RFID tags, operating at 868MHz, was deployed on the surface of pallets in a high-latitude power supply warehouse. The measured read / write distance was 2.8 meters at -20℃. Each tag stores pallet number information and can be associated with equipment coding. An InvenSense ICM-20948 nine-axis sensor was installed in the central recess of the pallet, acquiring acceleration (±8g, accuracy 0.02m / s²) at a sampling frequency of 200Hz. 2 The system incorporates data from a gyroscope (±1000° / s, accuracy 0.02° / s) and a magnetometer, adapting to a wide operating temperature range of -30°C to 50°C. Sixteen waterproof Bluetooth sensors are deployed on the warehouse ceiling, achieving clock synchronization via the IEEE 1588-2008 protocol, maintaining a synchronization accuracy of 300ns even at extreme temperatures of -30°C. Compressed sensing technology is used to acquire sensor data, which is then compressed using a Gaussian random measurement matrix before transmission and reconstructed using an orthogonal matching pursuit (OMP) algorithm, achieving a data compression ratio of 4:1 and effectively reducing communication power consumption by 35% in low-temperature environments.

[0052] 2. Graph-optimized multi-pallet cooperative localization model

[0053] When multiple pallets enter the positioning area, the relative distance between each pallet (accuracy ±3cm) is obtained through a UWB ranging module, and a relative position constraint equation is constructed. All pallet position variables are used to form a graph model, with relative distance constraints as edges, and a graph optimization algorithm is employed to solve for the globally optimal position. Innovatively, topological constraints of warehouse shelves (such as row and column alignment) are introduced, and the confidence level of each constraint is quantified through an information matrix to achieve collaborative optimization of multi-pallet positions. When simultaneously positioning 10 pallets, the parallel computing speedup reaches 4.2 times, the positioning time is reduced from 4.5s to 1.3s, and the multi-pallet discrimination rate is improved from 82% to 99%, solving the problem of multi-pallet confusion in traditional independent positioning schemes. In dense warehouse storage scenarios, this model can simultaneously handle the real-time positioning needs of more than 20 pallets, maintaining a positioning success rate of over 96%.

[0054] 3. Energy-accuracy balance strategy driven by particle swarm optimization

[0055] A Pareto optimization model for sensor energy consumption and positioning accuracy was established, with the total energy consumption model considering the combined effect of linear power consumption and nonlinear losses. A particle swarm optimization (PSO) algorithm was used to dynamically adjust the sampling frequency of each sensor: the inertia weight decreased linearly with the number of iterations, and the acceleration constant was adaptively adjusted to balance convergence speed and global optimization capability. In practical applications, the Bluetooth sensor sampling frequency was intelligently reduced from 100Hz to 60Hz, resulting in a 35% reduction in daily system energy consumption and an extension of battery life from 20 days to 32 days, while maintaining a positioning accuracy of 8.2cm. At -25℃, the system automatically increased the RFID reader power from 27dBm to 30dBm based on temperature sensor feedback to compensate for signal attenuation in low-temperature environments and ensure stable positioning performance.

[0056] 4. Sensor error compensation with enhanced low-temperature robustness

[0057] Design a deep learning-based environment adaptive model: A convolutional neural network is used to extract sensor features under low-temperature conditions, combined with LSTM to capture temporal dependencies, and the fusion weights are automatically adjusted. When a sudden temperature change is detected (such as temperature fluctuations caused by extreme weather), a robust Kalman filter is triggered, dynamically adjusting the robustness factor in the gain matrix to suppress the impact of temperature drift on the sensor. A sensor error compensation library for low-temperature environments is constructed, and parameters such as the zero bias and scale factor of the nine-axis sensor are calibrated for temperature. A polynomial compensation model is established, reducing the heading angle error of the magnetometer from ±15° to ±2.3° at -30°C. Simultaneously, a three-dimensional mapping table of temperature, energy consumption, and accuracy is constructed using a Gaussian mixture model to achieve automatic parameter adaptation under different temperature zones. A comparison of innovative technologies is shown in Table 2.

[0058] Table 2

[0059]

[0060] Multi-source heterogeneous data acquisition and spatiotemporal synchronization: In a smart logistics warehousing scenario, ThingMagicM6e UHF RFID tags are deployed in a 3×3 array on the pallet surface. Each tag stores static information such as pallet ID, material, and load-bearing capacity, operating at a frequency of 920-925MHz, with a reading distance of 3-5 meters. Bosch BNO055 six-axis sensors are installed at the four corners of the pallet to acquire three-axis acceleration (range ±16g) and three-axis gyroscope (range ±2000° / s) data at a frequency of 200Hz, with measurement accuracies of 0.01m / s². 2 And 0.01° / s. Sixteen Bluetooth ranging sensors are deployed at key locations in the warehouse to form a spatial monitoring network.

[0061] Sensor data synchronization is achieved through the IEEE 1588 precise time protocol, with a timestamp accuracy of 100ns. During data acquisition, the RFID reader polls the tags at 50ms intervals, and sensor data is written to a circular buffer according to timestamps, forming a spatiotemporally consistent data cube. When multiple tags respond simultaneously, the time-slotted ALOHA algorithm is used to avoid collisions, achieving a tag recognition success rate of 99.8%.

[0062] Multimodal feature fusion and dynamic weighted positioning: Feature-level fusion of collected multi-source data. RFID signal strength (RSSI) is calculated using the formula... Transform into a distance estimate, where P t For transmission power, G t G r Here, λ represents the gain of the transmitting and receiving antennas, λ is the wavelength, and d is the distance. After preprocessing with sliding window filtering (window size 10), the sensor data is fused using an improved DS evidence theory. The basic probability allocation function is calculated as follows:

[0063] Conflict coefficient

[0064] Introducing dynamic weighting factors Where λ = 0.5, the weights are adaptively adjusted based on the rate of change of sensor data. For example, when the tray moves rapidly, the weight of the inertial sensor increases to 0.7; when stationary, the weight of RFID is 0.8. Features are further fused through a deep learning model (3-layer fully connected network), and attention weights are added.

[0065] Multidimensional constraint localization solution: Constructing three-dimensional spatial constraint equations:

[0066]

[0067] Where (x,y,z) are the coordinates of the point to be located, (x,y,z) i ,y i ,z i Let θ be the coordinates of the reference point, wi be the weighting coefficient, di be the distance measurement value, and f(θ,φ,ψ) be the attitude constraint function, represented by quaternions. The Levenberg-Marquardt algorithm is used iteratively to solve the problem. The initial damping factor μ = 0.01 is adjusted in each iteration by μ ← μ × 10 (if the error increases) or μ ← μ / 10 (if the error decreases). The convergence threshold is 10. -6 .

[0068] In the positioning process, inertial navigation assistance is introduced, using formulas... Calculate the displacement increment and combine it with complementary filtering. Integrating inertial data and RFID data, among which

[0069] Dynamic environment perception and adaptive localization: Real-time classification of environment types is achieved using a random forest algorithm. The training dataset contains 5000 samples (1500 in open areas, 2000 in densely packed shelving areas, and 1500 in aisles), with a feature dimension of 20. In the densely packed shelving area, a reflected signal propagation model is introduced. Where ki is pre-calibrated using an environmental fingerprint database (ki = 0.6 for metal shelves, ki = 0.3 for wooden shelves), Δd i Estimated using a ray tracing algorithm.

[0070] An environmental fingerprint database was constructed using a Gaussian mixture model. K=5, and similarity is calculated using Kullback-Leibler divergence. When the similarity is below 0.8, adaptive adjustment of environmental parameters is triggered, including RFID signal thresholds and sensor fusion weights.

[0071] Anomaly Detection and Robustness Enhancement: A phase space reconstruction algorithm is designed to detect anomalies. Reconstruction parameters: embedding dimension m = 5, time delay τ = 10. The Euclidean distance between the reconstructed phase space trajectory and the standard trajectory is calculated. An alarm is triggered when the distance between 5 consecutive points exceeds a threshold ∈ = 1.5σ. A robust Kalman filter is employed, with a gain matrix K... k =P k|k-1 H T HP k|k-1 H T +R+λI) -1 λ = 0.1, suppressing the influence of outliers.

[0072] The timing anomaly scoring mechanism is passed through Calculate, N=10, w i =e -i / 5 When S t When the value is greater than 2.5, the data repair process is initiated, and outliers are replaced by regression predictions based on historical data.

[0073] Energy consumption optimization management and multi-tray collaborative positioning: establishing an energy consumption model The sensor sampling frequency was adjusted using a particle swarm optimization algorithm. Algorithm parameters: number of particles 20, number of iterations 50, inertia weight w = 0.7, acceleration constants c1 = c2 = 1.4. Experiments show that after dynamic adjustment, the system's daily energy consumption decreased by 32%, and battery life was extended by 45 days.

[0074] When multiple pallets are used for cooperative positioning, relative position constraint equations are constructed. Using the graph optimization algorithm min x ∑ i，j∈ε e ij (x) T Ωij e ij (x) Solution: When 5 pallets are positioned simultaneously, the average positioning time decreases from 3.2s to 1.8s, and the positioning success rate increases from 85% to 96%.

[0075] 7. Performance Evaluation and Data Analysis

[0076]

[0077] The above data demonstrates that the method presented in this application significantly outperforms traditional solutions in terms of positioning accuracy, environmental adaptability, robustness, and energy consumption optimization. The improved positioning accuracy is attributed to deep fusion of multimodal data and multidimensional constraint solving; enhanced adaptability to complex environments stems from dynamic environmental perception and adaptive parameter adjustment; improved anomaly handling capabilities rely on robust filtering and anomaly detection mechanisms; and reduced energy consumption is attributed to intelligent sampling frequency optimization. These technological innovations enable the system to operate efficiently and stably in practical applications, significantly improving the intelligence level and operational efficiency of logistics and warehousing management.

[0078] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A pallet detection and positioning method based on radio frequency identification and sensor fusion, characterized in that, Includes the following steps: 1) Multi-source heterogeneous data acquisition and spatiotemporal synchronization: An array of UHF RFID tags is installed on the surface of the pallet, with each tag storing feature information of different areas; a MEMS six-axis sensor is installed in the pallet; a Bluetooth ranging sensor array is installed in the warehouse; and time synchronization of all sensor data is achieved through the IEEE1588 time protocol to construct a spatiotemporally consistent data cube. 2) Multimodal feature fusion and dynamic weighted localization: Data fusion is performed using the DS evidence theory, through formulas... Calculate the basic probability distribution after fusion; where the conflict coefficient is... m1(A i )m1(A i ) and m2(A j Proposition A represents RFID and sensor data pairs, respectively. i and A j Basic probability allocation; introduction of dynamic weighting factors Where λ = 0.5 is the forgetting factor, x t The feature value of the sensor data at time t is used to adaptively adjust the weights of each mode based on the rate of change of the sensor data. 3) Multidimensional constraint localization solution: Constructing three-dimensional spatial constraint equations Where (x, y, z) are the three-dimensional coordinates of the pallet to be positioned, (x, y, z) i y i , z i Let be the three-dimensional coordinates of the i-th reference point, and w be the coordinates of the i-th reference point. i Let d be the weight coefficient for the i-th reference point. i Let f(θ, φ, ψ) be the measured distance from the pallet to be positioned to the i-th reference point; f(θ, φ, ψ) be the approximate pallet orientation d. i The beam functions, θ, φ, and ψ, represent the roll angle, pitch angle, and yaw angle of the pallet, respectively. The Levenberg-Marquardt algorithm is used to iteratively solve the nonlinear equations, with a convergence threshold set to 10. -6 .

2. The pallet detection and positioning method based on radio frequency identification and sensor fusion according to claim 1, characterized in that, Also includes: Dynamic environment perception and adaptive positioning steps: Real-time classification of environment types using machine learning algorithms, and dynamic adjustment of positioning parameters based on environmental characteristics; in the shelving area, an shelving reflection signal propagation model is introduced. Compensation for the multipath effect; where d eff The effective propagation distance is d, where d is the direct propagation distance and k is k. i Let Δd be the reflection coefficient of the i-th reflection path. i The additional path length for the i-th reflection path; Anomaly detection and robustness enhancement steps: An anomaly detection algorithm based on phase space reconstruction is designed. The attractor trajectory of the reconstructed phase space from sensor data is calculated. An anomaly handling mechanism is triggered when the trajectory deviation exceeds a threshold ∈ = 1.5σ. A robust Kalman filter algorithm is employed. in This is the optimal state estimate at time k. Let K be the predicted state at time k. k Let z be the gain matrix. k Let H be the observation value at time k, H be the observation matrix, and K be the gain matrix. k =P k|k-1 H T HP k|k-1 H T +R+λI) -1 λ is the robustness factor, P k|k-1 To predict the covariance matrix, R is the observation noise covariance matrix, and I is the identity matrix.

3. The pallet detection and positioning method based on radio frequency identification and sensor fusion according to claim 1, characterized in that, In step 1), compressed sensing technology is used to collect sensor data, and data compression is achieved using the formula y = Φx; where x is the original signal vector, Φ is the measurement matrix, and y is the compressed data vector; the reconstruction algorithm uses the Orthogonal Matching Pursuit (OMP) algorithm, with a number of iterations. μ = 1.2 is the overcompleteness coefficient, and σ0(x) is the sparsity of the original signal x.

4. The pallet detection and positioning method based on radio frequency identification and sensor fusion according to claim 1, characterized in that, In step 2), a deep learning model is used for feature fusion, and the importance weights of each modality are automatically learned through an attention mechanism, as shown in the formula: Among them, h i Let W1 and W2 be the feature vectors of the i-th mode, and W1 and W2 be the learnable weight matrices, α i Let be the attention weight for the i-th modality.

5. The pallet detection and positioning method based on radio frequency identification and sensor fusion according to claim 1, characterized in that, In step 3), inertial navigation-assisted positioning is introduced, using the formula... Calculate the displacement increment; where Δp is the displacement increment, v(t) is the velocity at time t, and a(τ) is the acceleration at time τ; inertial data and RFID data are fused using complementary filtering, and the fusion formula is as follows: in The position estimate after fusion at time t. For the position estimate at time t-1, To integrate weights, These are the variances of RFID and inertial measurement, respectively. The position data of the inertial measurement unit at time t.

6. The pallet detection and positioning method based on radio frequency identification and sensor fusion according to claim 1, characterized in that, In step 3), multi-frequency RFID signal fusion technology is used, through the formula... Calculate the distance; where d is the distance from the tray to be positioned to the RFID reader, c is the speed of light, f is the RFID signal operating frequency, and Phase is the RFID signal phase difference; construct a weighted least squares problem by combining multi-frequency measurements. Where x is the coordinate of the pallet to be positioned. The weight of the measurement value at the i-th frequency point. Let p be the variance of the measurement at the i-th frequency point, di be the distance measured at the i-th frequency point, and p be the variance of the measurement at the i-th frequency point. i Let be the coordinates of the i-th reference point.

7. The pallet detection and positioning method based on radio frequency identification and sensor fusion according to claim 2, characterized in that, In the dynamic environment perception and adaptive localization step, an environmental fingerprint database is constructed, and a Gaussian mixture model is used. Perform environment matching; where p(x) is the probability density function of environment feature x, K is the number of Gaussian distributions, and π k Let be the mixing coefficient of the k-th Gaussian distribution. The mean is μ k The covariance matrix is ​​∑ k Gaussian distribution; Kullback-Leibler divergence is used. Calculate the similarity between the current environment and the fingerprint database, where p(x) and q(x) are two probability distribution functions.

8. The pallet detection and positioning method based on radio frequency identification and sensor fusion according to claim 2, characterized in that, In the abnormal state detection and robustness enhancement steps, a time-series anomaly scoring mechanism is designed, using the formula... Calculate the outlier score; where S t Let w be the outlier score at time t, N be the sliding window size, and w be the outlier score at time t. i =e -i / τ The time decay weight is τ = 5, which is the time constant, and x is the time decay weight. t-i For the actual data at time ti, For the predicted data at time ti; when S t An alarm is triggered when the threshold is exceeded for three consecutive times.

9. The pallet detection and positioning method based on radio frequency identification and sensor fusion according to claim 1, characterized in that, Also includes: Energy consumption optimization management steps: Establish sensor energy consumption model Among them, E total For total energy consumption, E i Let be the energy consumption of the i-th sensor, and α = 0.01 be the nonlinear coefficient; a particle swarm optimization algorithm is used. Dynamically adjust the sensor sampling frequency; where Let be the velocity vector of particle i at time t+1, and w = 0.7 be the inertial weight. Let be the velocity vector of particle i at time t, c1 = c2 = 1.4 be the acceleration constant, r1 and r2 be random numbers between [0, 1], and p i Let be the optimal position of particle i. Let p be the position of particle i at time t. g This is the globally optimal position.

10. The pallet detection and positioning method based on radio frequency identification and sensor fusion according to claim 1, characterized in that, Also includes: Multi-pallet cooperative positioning steps: When each pallet enters the positioning area, construct the relative position constraint equations between the pallets. Where p i p j Let d be the coordinate vectors of tray i and tray j, respectively. ij The relative distance measurement between tray i and tray j, ∈ ij To mitigate measurement error, a graph optimization algorithm (min) is employed. x ∑ i，j∈ε e ij (x) T Ω ij e ij (x) Solve for the global optimal position; where x is the set of position variables for all pallets, ε is the edge set of relative position constraints between pallets, and e ij (x) is the error function of the relative positions of tray i and tray j, Ω ij This is an information matrix.

Images